Release 0.20

Release 0.20

This page records the notes of the 0.20 release, with its patches up to 0.20.8.

Released 2026-10-04 to 2026-10-05. This release is a full reset on Claude Code 2.1.289. It cut runtime JavaScript from 18,935 lines to 2,118 lines. It adds a new status line, the dotclaude-jev plugin, the Terms of Use, and the CodeGraph augment. A version in parentheses marks a later patch of this line. An entry with no version comes from 0.20.0.

0.20.0 keeps the parts that serve a need that Claude Code does not cover. It removes the rest. Runtime JavaScript goes from 18,935 lines in 0.19.1 to 2,118 lines. tests/budget.test.mjs bounds it at 3,000. Run /dotclaude:setup again after you update.

Added

Reset Parts

• docs/parts.md, a page that lists each part, its event, its bound, and the need that it serves.
• The rule that runtime JavaScript (hooks/, status-line/, skills/, plugins/) stays within RUNTIME_JS_LINES (3,000) in hooks/lib/_budget.mjs. 0.19.1 had no bound on it. Its runtime JavaScript grew to 18,935 lines. Much of it was a copy of what Claude Code 2.1.287 to 2.1.289 does itself. 0.20.8 removed this bound (see Changed, 0.20.8).
• The plugin options guard_agents and compaction_handoff. guard_agents turns off the spawn rules. compaction_handoff turns off the handoff fork.
• A Stop check gate. It sends Claude back once when a turn edited files and no check ran after the last edit. The usage evidence shows that a rule in a prompt does not hold, and a hook does. It finds the check from the justfile, package.json, or Makefile of the project.
• A compaction instruction. The summary says that an unapproved plan or an open question stays open. The default summary text tells Claude to continue without asking. A model then acted on a plan that the user had not approved.
• A handoff fork on compaction. Before a compaction of the main conversation, a fork writes a handoff note to .claude/handoffs/. It adds the note to the compacted conversation, because users report that a handoff note beats a compaction. If the fork gives no note in 60 seconds, the compaction runs without it.
• Cold-cache notes. A resumed session with an expired prompt cache gets the cost advice. A prompt that comes more than 1 hour after the last turn gets it too. The reason is that 1.6% of the turns caused 80% of the cache writes. Those are the turns after more than 1 hour idle.
• Plan detection from the account in .claude.json, in hooks/lib/_plan.mjs. The cold-cache note uses a 5 minute cache time on the api plan. It uses 1 hour on other plans. A new or cleared session on the api plan gets a plan note. skills/setup/scripts/settings.mjs shows the plan. It takes --plan <id>, and it applies the per-plan plans overrides of the profile.
• The backend option of dotclaude-browser, which defaults to agent-browser. The value cloakbrowser runs agent-browser with the CloakBrowser binary through --executable-path. Use it for sites with bot detection.

Status Line

• A status line with every 0.19 part, in less code. Each part has one shape: icon, bar, number.
  • One bar renderer, one icon set, and one color scale (USAGE_LEVELS) serve every percentage.
  • The rows are core (model, context, cache, limits), place (folder, branch, worktree, PR, loop), and detail (cache hit ratio and misses, limit pace, cost).
  • A ⚠ blinks for three cases. The first is a handoff due, at context 90% of the compaction point. The second is a cache that expires in two minutes or less. The third is a limit at 90%.
  • ⚠ also marks an effort that the rules do not allow for the model.
  • A second script shows the model and context of each running agent.
  • /dotclaude:setup writes two stubs in the config directory. It sets statusLine and subagentStatusLine to run them. A status line command gets an empty ${CLAUDE_PLUGIN_ROOT}.
• Limit resets (0.20.1): the status line shows the limit resets that you can use, as ⟳2 by Oct 12 at 4am. That text gives the count and the earliest expiry.
  • It reads them from the cedar_ember block of the /usage copy that Claude Code keeps in .claude.json. So it makes no network call.
  • A grant counts when it is not paused, has resets left, and is in its start to end span. Claude Code and CodexBar count a grant the same way.
  • While a usage window is at its limit, the reset part outranks the other parts.
• Extra usage (0.20.1): the status line shows the extra usage spend while extra usage is on. An example is extra █░░░░ $12.50/$50. With no monthly limit, it shows extra $12.50. It reads extra_usage from the same copy.

Setup and Profile (0.20.4)

• The profile sets cleanupPeriodDays to 14. It sets enableArtifact to false. It sets disableBundledSkills to true. It sets CLAUDE_CODE_DISABLE_EXPLORE_PLAN_AGENTS to 1.
• /dotclaude:setup keeps the newest 3 backups of each file. It removes the older ones. It lists the memory files to review. It lists the LSP plugins for the language servers on PATH.
• The setup integrations reference covers the LSP plugins and the ctx7 CLI for library docs. It has no MCP server. The web-researcher agent uses ctx7 when it is on PATH, because it uses less context than a web search. When ctx7 is rate-limited, the agent checks the limit with one call. Then it uses WebSearch and WebFetch until the reset.
• /dotclaude:setup gives the command that adds a Context7 API key to the shell profile. The user runs it outside Claude Code. Setup then checks that the key gives a tier other than anonymous.
• The README has a "Save weekly usage" section. The dossier ranks the community claims about usage.

Rules, Notes, and Hooks (0.20.4)

• Each compaction note marks the earlier compaction note of the same session as superseded. The new note carries its open items. Before, each compaction left one more in-progress note.
• The CodeGraph augment: when a Grep call or a Bash rg or grep search looks for one symbol name in a project with .codegraph/. The hooks module adds the callers and callees of that symbol to the search result.
  • Claude ignored rules that told it to run codegraph. So the graph now comes with the search that Claude already runs.
  • The design follows the GitNexus hooks.
  • Before a lookup, the hook runs codegraph sync if files changed since the last index. Without the MCP server, the index does not update itself.
  • It never builds an index.
  • The plugin option codegraph turns it off.
• The minimal-code rules: session start tells Claude to build the minimum, with the idea of Ponytail. The plugin option ponytail turns them off.
• Usage habits in the working rules: a handoff note and /clear at a task boundary. /btw for a side question. A lower effort in place of a model change. A model change only after /clear.
• Line breaks in the working rules: in Markdown, comments, commits, and PRs, Claude starts each sentence on a new line. Claude does not break lines at a column. The rule follows Semantic Line Breaks.
  • Claude hard-wrapped prose at about 80 columns. That breaks reflow in editors and terminals, and it makes diffs larger.
  • A project style takes precedence.
  • Shorter wording of the other rules keeps the file within its 2,000-byte bound.
  • Claude Code 2.1.289 has no CLAUDE_NO_WRAP variable or --no-word-wrap flag. A plugin cannot change how the terminal wraps output.

Jev Plugin

• The optional dotclaude-jev plugin (0.20.4): its second-opinion skill asks TypeSafe Jev, a decision model. It asks for a yes or no check, a pick from options, or a rating. Each answer has a probability.
  • Claude uses it for its own close calls. It still asks the user the questions that only the user can answer.
  • It also checks a technical decision of the user. When Jev finds that facts decide it and that it causes defects, Claude tells the user the facts once. Claude asks for a confirmation. The answer of the user is final.
  • It needs TYPESAFE_API_KEY.
• The dotclaude-jev plugin (0.20.8) adds a session note that sends each decision through the second-opinion skill.
  • The decisions are the close calls of Claude. They are the questions of the user that ask for a decision, a pick, or a rating. They are the technical decisions of the user, and each request for a second opinion.
  • Claude asks each question with options through AskUserQuestion. So the hooks module can add the pick of Jev.
  • Goals, preferences, and approvals stay with the user.

Removed

• The delegation gate and note. Users report cost blow-ups from subagents. maxTurns bounds an agent without it.
• The agent budget hooks (enforce-agent-budget, hand-off-capped-agents, and the report caps). The maxTurns of an agent file does the same job.
• The model-switch, ConfigChange, StopFailure, and TaskCompleted hooks. availableModels and CLAUDE_CODE_DISABLE_FAST_MODE do the same job.
• The options model_lock, model_plan, model_allowed, and usage_notes. The other 0.19 options go too, except the guard options. The hooks that they controlled are gone.
• Fable 5.1 from availableModels and from the Agent(model:fable*) deny rule. The user chose to remove it. A subagent never runs on it.
• Nested instructions (load-nested-instructions). Claude Code 2.1.288 loads rules on Write and Edit. A Bash read of a directory does not load its CLAUDE.md yet.
• The restore after a compaction, the check for line breaks, and the scratchpad pruning. The 0.20.0 review found no need for them that Claude Code does not cover.
• warn-instruction-size. /doctor audits the instruction files.
• The skills slices, explain, and polish. /goal covers a large change. The 0.20.0 review found no need for the other two.
• The agent mechanical-worker. implementer does its work.
• The output styles Explanatory, Learning, and Proactive. Claude Code has built-in styles of those names.
• The recognize-captcha skill and the CloakBrowser launcher. CloakBrowser keeps most CAPTCHAs away. agent-browser --executable-path runs its binary.
• The AI policy catalog and its update script scripts/update-ai-policies.mjs. scripts/count-tokens.mjs goes too. They served features that 0.20.0 removed. The contribute skill still reads the policy of the project.
• The eval case t4-slices, because it tests the removed slices skill.
• The 0.19 profile optional.json, the setup scripts for migration and the managed lock, and the public docs pages for the removed parts.

Related Pages

• Previous: Release 0.19
• Next: Release 0.21
• Changes and fixes: Release 0.20 Changes and Fixes
• Archive: the index of the archived pages.